AI Roundup 056: Data deals
March 1, 2024.
Housekeeping
We’re trying some new things this week! First, Substack just launched some new DM features. If you’re a paid subscriber, send me your questions for an upcoming Q&A section (or just say hi!).
Second, there’s a new section: “Fundraising.” In researching the week’s news, I come across a lot of funding announcements for AI startups - they’re included as a bonus for paid subscribers at the bottom of the roundup. Thanks for reading!
Data deals
Following on the heels of Reddit's announcement last week, Stack Overflow is also partnering with Google to license its data. Google plans to use the new data API to enrich Gemini.
Why it matters:
Last year, user-generated content companies discovered the value of their posts and comments and quickly put up new restrictions. Now they're offering said content again, through a licensing regime.
In addition to Stack Overflow, Tumblr and Wordpress are reportedly preparing to sell UGC to Midjourney and OpenAI.
It's hard to blame them: Reddit has already made $203 million from data licensing (to be paid out over 2-3 years).
Elsewhere in the FAANG free-for-all:
After a decade in development, Apple has killed its electric/autonomous car project, and is moving some employees to generative AI initiatives. Meanwhile, Tim Cook said Apple will break new ground in AI this year.
Google is reportedly paying a handful of publishers to use unreleased AI tools to write stories - publishers are expected to produce three articles per day, one newsletter per week, and one marketing campaign per month with the tools.
And Microsoft launches Copilot for Finance in public preview, helping users reconcile data in Excel and speed up the collections process in Outlook.
Sundar gets serious
Despite Google's best efforts, the backlash to Gemini's "woke" responses grew this week. It culminated in the company's CEO, Sundar Pichai, calling the diversity errors "completely unacceptable" in an internal memo.
Between the lines:
People were really upset about this. Not just conservatives decrying Google's hidden woke agenda, but also liberals and techies who see it as the canary in the coal mine of Google's demise (and are calling for Sundar's resignation).
Google employees have described the issue as "a technical fix to reduce bias" that backfired, though others are using it as a cautionary tale as to why AI can't be safely controlled.
Maybe Google is on a downward trajectory, but its multi-billion-dollar money printer will keep it alive for a long time. And while there are real AI safety and transparency concerns here, Gemini's bias snafu is a new or unique class of problem.
So, the hysteria seems a bit silly. To quote Max Read: "imagine getting so mad at your computer because it won’t say whether Elon Musk or Hitler is worse that you insist that the head of the computer company needs to step down."
Elsewhere in AI anxiety:
At least 100 cases of malicious ML models were found on Hugging Face, some of which can execute code on users' machines.
"BadGPT" and "FraudGPT" are two examples of LLMs sold on the dark web to write phishing emails, create fake websites, and create malware.
A look at how AI is casting a long shadow on the adult entertainment industry, as AI "dream girls" threaten to replace human actresses.
And OpenAI faces two new lawsuits: one from publications over copyright infringement and one from Elon Musk over abandoning its mission.
IA française
If you don't pay close attention to open-source releases, you might not know much about Mistral - a 9-month-old French AI startup valued at over $2 billion.
In the news:
The company has already made a big name by releasing small(er), open-source models that rival some of the best proprietary ones.
Now, the company is also releasing Mistral Large (notably not open-source), a GPT-4 rival, and Le Chat, an assistant to compete with ChatGPT.
And the company has even secured a (small) investment from Microsoft, though the deal is already facing scrutiny from EU watchdogs.
Elsewhere in model releases:
Meta plans to launch Llama 3 this July and wants it to be comfortable answering more controversial questions, according to The Information.
Qualcomm unveiled the AI Hub, a library of over 75 pre-optimized AI models that can be deployed on mobile devices powered by Snapdragon and Qualcomm platforms.
Hugging Face, ServiceNow, and Nvidia released StarCoder 2, a code-generating AI that can run on most consumer GPUs.
And an honorable mention: Alibaba researchers detailed Emote Portrait Alive (EMO), a new framework that can create a talking or singing head from a still image and an audio file.
Things happen
Klarna says its AI-powered support bot has replaced 700 human agents. Brave debuts new AI assistant Leo for Android users. The UK government is piloting in-house AI to improve efficiency. Defense official says Project Maven algorithms helped tailor recent airstrikes. SK Telecom to partner with Perplexity and offer its search engine for free. AI-generated Kara Swisher biographies are flooding Amazon. Meta outlines strategy to combat AI misuse ahead of the June EU elections. Robotics startup Figure AI is partnering with OpenAI on models for humanoid robots. Nvidia's RTX 500 and RTX 1000, new laptop GPUs for on-the-go inference. Democratic consultant working for Biden challenger admits to deepfake robocalls. Arc launches a "pinch-to-summarize" AI feature in its app. Wikipedia downgrades CNET's reliability rating after AI-generated articles. Google Genie creates AI outputs resembling 2D video games. Man running AI-powered porn site horrified by what users are asking for. Here lies the internet, murdered by generative AI. OpenAI argues the NY Times paid someone to hack their products. AI deepfakes are cheap, easy, and coming for 2024 elections. FlowGPT is the Wild West of generative AI apps. Ghost kitchens are advertising AI-generated food on DoorDash.
